In this thesis work, we propose a reduced complexity implementation method of the Qualcomm CELP (QCELP) speech coding algorithm which is the standard for CDMA digital cellular system, and compare its performance with those of the VSELP and the RPE-LTP speech coders which are the standards for TDMA and GSM digital cellular systems, respectively.
The pitch period is usually quite continuous with time during the voiced segments. However, a frame-by-frame pitch search algorithm alone sometimes does not produce a smooth pitch contour since it often results in doubling or halving of the correct pitch period. So we need a pitch smoothing method.
For the pitch contour smoothing, we calculate variations of previous pitch periods. If this variations are smaller than a certain value, we consider that current speech segment is in a voiced region and then employ a proper weighting factor to decide the pitch period of current speech subframe. In this case, we can reduce the computational complexity by narrowing the search range for the current pitch period around the previous one. Experiments show that we can reduce the complexity of calculation in pitch search to 85% without degradation of speech quality.
We perform intelligibility and naturalness tests for the coded output speech obtained from each of QCELP, VSELP, and RPE-LTP coding algorithms. To obtain the two subjective measures, we use the Modified Rhyme Test (MRT) and Mean Opinion Score (MOS). We also obtain the performance of the respective coding algorithm for telephone conferencing by mixing coded voices of two speakers. It is measured by the sentence discrimination capability. The MOS test results show that the naturalness of QCELP at around 5Kbps is comparable to that of VSELP at 8Kbps and is better than that of RPE-LTP at 13Kbps. From the MRT, RPE-LTP yields the best performance while VESLP shows the worst intelligibility. The performance test on mixed sentences uttered by 2 different speakers gives sentence discrimination rate of 84% for VSELP and 73% for QCELP.